Hugging Face's logo

RichardErkhov
/
nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf

GGUF
Inference Endpoints
Model card Files Community
RichardErkhov commited on
Commit
bc17dbb
•
1 Parent(s): 1790c44

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +149 -0
README.md ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ cpt_st-vicuna-v1.3-1.5b-ppl - GGUF
11
+ - Model creator: https://huggingface.co/nota-ai/
12
+ - Original model: https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-1.5b-ppl/
13
+
14
+
15
+ | Name | Quant method | Size |
16
+ | ---- | ---- | ---- |
17
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q2_K.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q2_K.gguf) | Q2_K | 0.56GB |
18
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_XS.gguf) | IQ3_XS | 0.61GB |
19
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_S.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_S.gguf) | IQ3_S | 0.64GB |
20
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_S.gguf) | Q3_K_S | 0.64GB |
21
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_M.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.IQ3_M.gguf) | IQ3_M | 0.66GB |
22
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K.gguf) | Q3_K | 0.7GB |
23
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_M.gguf) | Q3_K_M | 0.7GB |
24
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q3_K_L.gguf) | Q3_K_L | 0.75GB |
25
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.IQ4_XS.gguf) | IQ4_XS | 0.77GB |
26
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q4_0.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q4_0.gguf) | Q4_0 | 0.81GB |
27
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.IQ4_NL.gguf) | IQ4_NL | 0.81GB |
28
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K_S.gguf) | Q4_K_S | 0.81GB |
29
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K.gguf) | Q4_K | 0.84GB |
30
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q4_K_M.gguf) | Q4_K_M | 0.84GB |
31
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q4_1.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q4_1.gguf) | Q4_1 | 0.88GB |
32
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q5_0.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q5_0.gguf) | Q5_0 | 0.96GB |
33
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K_S.gguf) | Q5_K_S | 0.96GB |
34
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K.gguf) | Q5_K | 0.98GB |
35
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q5_K_M.gguf) | Q5_K_M | 0.98GB |
36
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q5_1.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q5_1.gguf) | Q5_1 | 1.04GB |
37
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q6_K.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q6_K.gguf) | Q6_K | 1.13GB |
38
+ | [cpt_st-vicuna-v1.3-1.5b-ppl.Q8_0.gguf](https://huggingface.co/RichardErkhov/nota-ai_-_cpt_st-vicuna-v1.3-1.5b-ppl-gguf/blob/main/cpt_st-vicuna-v1.3-1.5b-ppl.Q8_0.gguf) | Q8_0 | 1.46GB |
39
+
40
+
41
+
42
+
43
+ Original model description:
44
+ # Shortened LLM Model Card
45
+
46
+ Shortened LLM is a depth-pruned version of large language models for efficient text generation.
47
+
48
+ - **Developed by:** [Nota AI](https://www.nota.ai/)
49
+ - **License:** Non-commercial license
50
+ - **Repository:** https://github.com/Nota-NetsPresso/shortened-llm
51
+ - **Paper:** https://arxiv.org/abs/2402.02834
52
+
53
+ ## Compression Method
54
+ * After identifying unimportant Transformer blocks, we perform **one-shot pruning**.
55
+ * In retraining pruned models for quality recovery, **continued pretraining (CPT)** on a large corpus markedly outperforms LoRA-based tuning, particularly at severe pruning ratios.
56
+
57
+ ## Models from Aggressive Pruning & CPT Retraining (arXiv-v2):
58
+ | Source<br>Model | Pruning<br>Ratio | Pruning<br>Criterion | HF Models<br>Link |
59
+ |:---:|:---:|:---:|:---:|
60
+ | Vicuna-v1.3-7B | 20% | PPL | [nota-ai/cpt_st-vicuna-v1.3-5.5b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-5.5b-ppl) |
61
+ | Vicuna-v1.3-7B | 45% | PPL | [nota-ai/cpt_st-vicuna-v1.3-3.7b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-3.7b-ppl) |
62
+ | Vicuna-v1.3-7B | 60% | PPL | [nota-ai/cpt_st-vicuna-v1.3-2.7b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-2.7b-ppl) |
63
+ | Vicuna-v1.3-7B | 80% | PPL | [nota-ai/cpt_st-vicuna-v1.3-1.5b-ppl](https://huggingface.co/nota-ai/cpt_st-vicuna-v1.3-1.5b-ppl) |
64
+
65
+ <details>
66
+ <summary>
67
+ Click to see the results:
68
+ </summary>
69
+
70
+ - EleutherAI/lm-evaluation-harness version [3326c54](https://github.com/EleutherAI/lm-evaluation-harness/tree/3326c547a733d598b4377e54be96e194861b964c)
71
+
72
+ <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st_llm-cpt_results.png" width="100%">
73
+
74
+ </details>
75
+
76
+ #### Experimental Setup for CPT of Pruned Vicuna-7B
77
+ * Dataset: [SlimPajama-627B](https://huggingface.co/datasets/cerebras/SlimPajama-627B)
78
+ * Training using 8 NVIDIA H100 GPUs.
79
+ * 5.5B parameters: 37B training tokens (for 6 days)
80
+ * 3.7B parameters: 74B tokens (for 8 days)
81
+ * 2.7B parameters: 150B tokens (for 12 days)
82
+ * 1.5B parameters: 271B tokens (for 11 days)
83
+ * AdamW optimizer with (β1, β2)=(0.9, 0.95); a learning rate of 0.0001; a weight decay of 0.1.
84
+ * Global batch size: 512 (micro-batch size of 2 × 32 gradient accumulation steps × 8 GPUs).
85
+
86
+ <details>
87
+ <summary>
88
+ Click to see the learning curve:
89
+ </summary>
90
+
91
+ **Zero-shot performance over the course of training for models from Vicuna-7B-v1.3 at different pruning ratios.** For each model size, the CPT duration was limited to a two-week period, but additional training could further improve the quality.
92
+
93
+ <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st_llm-cpt_learning-curve.png" width="100%">
94
+
95
+ </details>
96
+
97
+
98
+
99
+ ## Models from Moderate Pruning & LoRA Retraining (arXiv-v1):
100
+ | Source<br>Model | Pruning<br>Ratio | Pruning<br>Criterion | HF Models<br>Link |
101
+ |:---:|:---:|:---:|:---:|
102
+ | LLaMA-1-7B | 20% | PPL | [nota-ai/st-llama-1-5.5b-ppl](https://huggingface.co/nota-ai/st-llama-1-5.5b-ppl) |
103
+ | LLaMA-1-7B | 20% | Taylor+ | [nota-ai/st-llama-1-5.5b-taylor](https://huggingface.co/nota-ai/st-llama-1-5.5b-taylor) |
104
+ | Vicuna-v1.3-7B | 20% | PPL | [nota-ai/st-vicuna-v1.3-5.5b-ppl](https://huggingface.co/nota-ai/st-vicuna-v1.3-5.5b-ppl) |
105
+ | Vicuna-v1.3-7B | 20% | Taylor+ | [nota-ai/st-vicuna-v1.3-5.5b-taylor](https://huggingface.co/nota-ai/st-vicuna-v1.3-5.5b-taylor) |
106
+ | Vicuna-v1.3-13B | 21% | PPL | [nota-ai/st-vicuna-v1.3-10.5b-ppl](https://huggingface.co/nota-ai/st-vicuna-v1.3-10.5b-ppl) |
107
+ | Vicuna-v1.3-13B | 21% | Taylor+ | [nota-ai/st-vicuna-v1.3-10.5b-taylor](https://huggingface.co/nota-ai/st-vicuna-v1.3-10.5b-taylor) |
108
+
109
+ <details>
110
+
111
+ <summary>
112
+ Click to see the results:
113
+ </summary>
114
+
115
+ - EleutherAI/lm-evaluation-harness version [3326c54](https://github.com/EleutherAI/lm-evaluation-harness/tree/3326c547a733d598b4377e54be96e194861b964c)
116
+
117
+ <img alt="results" img src="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/compressed-llm/st-llama_zero-shot_scores.png" width="100%">
118
+
119
+ </details>
120
+
121
+ ## License
122
+ - All rights related to this repository and the compressed models are reserved by Nota Inc.
123
+ - The intended use is strictly limited to research and non-commercial projects.
124
+
125
+ ## Acknowledgments
126
+ - [Microsoft for Startups Founders Hub](https://www.microsoft.com/en-us/startups) and [Gwangju AICA](http://www.aica-gj.kr/main.php) for generously providing GPU resources.
127
+ - [LLM-Pruner](https://github.com/horseee/LLM-Pruner), which utilizes [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness), [PEFT](https://github.com/huggingface/peft), and [Alpaca-LoRA](https://github.com/tloen/alpaca-lora). Thanks for the pioneering work on structured pruning of LLMs!
128
+ - Meta AI's [LLaMA](https://github.com/facebookresearch/llama) and LMSYS Org's [Vicuna](https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md). Thanks for the open-source LLMs!
129
+
130
+ ## Citation
131
+ ```bibtex
132
+ @article{kim2024shortened,
133
+ title={Shortened LLaMA: Depth Pruning for Large Language Models with Comparison of Retraining Methods},
134
+ author={Kim, Bo-Kyeong and Kim, Geonmin and Kim, Tae-Ho and Castells, Thibault and Choi, Shinkook and Shin, Junho and Song, Hyoung-Kyu},
135
+ journal={arXiv preprint arXiv:2402.02834},
136
+ year={2024},
137
+ url={https://arxiv.org/abs/2402.02834}
138
+ }
139
+ ```
140
+ ```bibtex
141
+ @article{kim2024mefomo,
142
+ title={Shortened LLaMA: A Simple Depth Pruning for Large Language Models},
143
+ author={Kim, Bo-Kyeong and Kim, Geonmin and Kim, Tae-Ho and Castells, Thibault and Choi, Shinkook and Shin, Junho and Song, Hyoung-Kyu},
144
+ journal={ICLR Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo)},
145
+ year={2024},
146
+ url={https://openreview.net/forum?id=18VGxuOdpu}
147
+ }
148
+ ```
149
+